KoboldCpp runs on one PC, and you can use it from other devices. KoboldAI Lite works in any browser, including on phones.
| Where the other device is | Use |
|---|---|
| Same Wi-Fi or network | The PC's network address |
| Anywhere | Remote Tunnel |
Same network
Section titled “Same network”KoboldCpp listens on all network interfaces by default, so devices on your network can already reach it.
- If others share your network, set a Password: on the Network tab. Leave Host: on the same tab empty; with
127.0.0.1, only this PC can connect. - Launch KoboldCpp.
- Find the PC's IP address on your network, for example
192.168.1.20.- Windows: run
ipconfigin Command Prompt and look for IPv4 Address. - Linux: run
hostname -I. - macOS: run
ipconfig getifaddr en0.
- Windows: run
- On the other device, open
http://<IP address>:5001, for examplehttp://192.168.1.20:5001.
The console still prints http://localhost:5001. That address is only for the PC itself.
If the page does not load, the PC's firewall is usually blocking port 5001, or the IP address is wrong:
- Windows: on first launch, Windows may ask whether KoboldCpp may use the network. Allow Private networks. If you missed it, open Windows Security > Firewall & network protection > Allow an app through firewall and tick KoboldCpp for Private.
- Linux (ufw):
sudo ufw allow 5001/tcp.
Browsers allow the microphone and camera only on HTTPS or on the PC itself (localhost). For voice input or the camera from another device, use the Remote Tunnel link or HTTPS.
For other apps, use the same address in place of localhost, for example http://192.168.1.20:5001/v1 for OpenAI-compatible apps.
Anywhere: Remote Tunnel
Section titled “Anywhere: Remote Tunnel”Remote Tunnel creates a public HTTPS link through Cloudflare. It works even behind a firewall.
Set a Password: on the Network tab.
Tick Remote Tunnel on the Quick Launch or Network tab. On the command line, add
--remotetunnel.Launch. The first time, KoboldCpp downloads Cloudflare's
cloudflaredprogram from GitHub.Wait for the link in the console:
Console Your remote tunnel is ready, please connect to https://<random-words>.trycloudflare.comOpen the link on any device to use KoboldAI Lite. For other apps, add
/apior/v1to it, as printed in the console.
The link changes every time you launch KoboldCpp.
Anyone who has the link can reach your KoboldCpp. The password keeps them out of text generation, but not out of image generation. See what the password protects.
If the tunnel fails, the console prints Error: Could not create cloudflare tunnel! or Remote Tunnel Failed!. KoboldCpp needs internet access to download cloudflared and to open the tunnel.
Remote Tunnel does not start in terminal-only modes (--cli, --prompt, --benchmark).
Share with the world: AI Horde
Section titled “Share with the world: AI Horde”To let anyone use your model through the AI Horde network instead of a private link, run KoboldCpp as a Horde worker. See Horde worker.
Advanced: choose the network address
Section titled “Advanced: choose the network address”Host: (--host) sets the address KoboldCpp listens on. Leave it empty for all interfaces, use 127.0.0.1 for this PC only, or one of the PC's IP addresses to listen on that network only. When Host: is set, the console prints that address in place of localhost.
The port is Port: on the Network tab (--port, default 5001).
Advanced: HTTPS on your network
Section titled “Advanced: HTTPS on your network”On your own network, KoboldCpp serves plain HTTP. To serve HTTPS, set SSL Cert: and SSL Key: (--ssl cert.pem key.pem). The console addresses then start with https://. See HTTPS (SSL).
With SSL on, Remote Tunnel connects to KoboldCpp over HTTPS and accepts its certificate even if it is self-signed.
Safety checklist
Section titled “Safety checklist”- Set a Password: before you use Remote Tunnel or share on your network.
- If admin mode is on, set an Admin Password:.
- Image generation and
/tts_to_audioare open to anyone who can reach the server. Load those models only on instances you trust, or limit image size with--sdclamped. - For shared servers, consider
--ratelimit,--genlimitand--maxrequestsize. See Limits for shared servers.